Semantic Primes Explain LLM Emotions More Effectively.

Understanding how large language models (LLMs) process and represent emotions is a growing area of research, but what constitutes a good explanation for these emotional mechanisms remains unclear. While emotion representations and circuits can be identified, they often lead to circular explanations or arbitrary dimensions. This study explores whether a more fundamental set of internal variables, specifically the "semantic primes" derived from the Natural Semantic Metalanguage (NSM), could offer superior explanations. Across four instruction-tuned LLMs (Llama-1B, Gemma-2B, Gemma-9B, OLMo-7B), experiments yielded compelling evidence. Firstly, NSM primes were found to be recoverable internal elements within these models. Secondly, intervening with a prime-based direction in a reference model demonstrated approximately three times stronger and twice as selective control over emotion compared to the best appraisal-based directions. Finally, the models themselves treated a prime-based explication as interchangeable with the corresponding emotion. These findings collectively suggest that NSM primes serve as more effective "explanans" for emotion in LLMs, aligning better with scientific explanation criteria than many alternative approaches. This offers a more primitive and potentially universal way to understand and control emotional responses in AI.